VLA GalleryAll modelsGallery

VLA Gallery / All models / Industry models

Gemini Robotics 1.5

Google DeepMind · 2025-09

VLAIndustry modelsNot disclosedNot disclosed

A 'Thinking VLA' that reasons in natural language before acting, training on data from multiple embodiments together via Motion Transfer.

View architecture diagram in the galleryPaper ↗

Paper

Gemini Robotics 1.5: Pushing the Frontier of Generalist Robots with Advanced Embodied Reasoning, Thinking, and Motion TransferarXiv 2510.03342

Overview

Category
Industry models
Type
VLA
Released
2025-09
Organization
Google DeepMind
Open weights
No
Parameters
Not disclosed

Architecture

Backbone LLM
Based on Gemini (size not disclosed)
Vision encoder
—
Action head
Interleaved generation of thinking·actions (Thinking VLA) (Not disclosed)
Chunk
Not disclosed
Control rate
Not disclosed
Architecture (original summary)
Gemini-based Thinking VLA (interleaves natural-language thinking and actions); multi-embodiment learning via Motion Transfer; action chunk and control frequency not disclosed

Training

Pretraining accelerators
TPU v4, v5p, v6e (chip count not disclosed; no split between pre- and post-training)
Pretraining duration
Not disclosed
Pretraining steps & batch
Not disclosed
GPU/TPU-hours
Not disclosed
Fine-tuning
Not disclosed
Data type
Real-robot action data (multiple embodiments) + web vision-language (public text·images·video)
Data scale
'thousands of diverse tasks'; time and episode count not disclosed
Main datasets
In-house multi-embodiment robot data (private) + public internet text, images, and video
Tasks & evaluation
230-task benchmark (visual · instruction · action · task generalization); bi-arm Franka 44 tasks (20 in-distribution + 24); ALOHA 105 tasks†; over 90% of evaluation episodes during development run in simulation; agentic comparison (Gemini 2.5 Flash + GR 1.5)
Robot
ALOHA 2, bi-arm Franka, Apptronik Apollo humanoid

Inference

Inference device & speed
Not disclosed
Inference memory
Not disclosed
Weights bf16 (calc.)
Not disclosed
Where it runs
Not disclosed

Reinforcement learning & base

Reinforcement learning
Not used · Technical report mentions RL only as future work
RL evidence
https://arxiv.org/html/2510.03342
Base control
Fixed · No mention of base·locomotion control (ALOHA·bi-arm Franka·Apollo manipulation tasks)

Notes

Key features
A 'Thinking VLA' that reasons in natural language before acting, training on data from multiple embodiments together via Motion Transfer. Reports that tasks trained only on ALOHA 2 also work on Apollo and bi-arm Franka.
Notes
Blog 2025-09-25, arXiv v1 2025-10-02 (based on the earlier blog). 'currently available to select partners'. The model card covers GR 1.5 and ER 1.5 together (Software: JAX, ML Pathways). ALOHA 105 tasks†: the confirmed sentences state only existing generalization 68 + new action generalization 5 + new task categories 12.
Verification
Confirmed

Other models in this category